enhancing performance through pedagogy and feedback: domain considerations for intelligent tutoring...
TRANSCRIPT
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 1 of 12
Enhancing Performance through Pedagogy and Feedback: Domain
Considerations for ITSs
Benjamin S. Goldberg, Heather K. Holden, Keith W. Brawner, and Robert A. Sottilare
U.S. Army Research Laboratory – Human Research and Engineering Directorate
Orlando, Florida
{benjamin.s.goldberg; heather.k.holden; keith.brawner; robert.sottilare }@us.army.mil
ABSTRACT
As computer-based instruction evolves to support more adaptive training, it is becoming increasingly more evident
that such programs be designed around an individual trainee's characteristics, rather than focusing just on task
performance. In other words, a trainee‟s state (e.g. how they learn, their affect and motivation) is an important factor
in performance and retention. To optimize individual performance in computer-based training Intelligent Tutoring
System (ITS) technologies (tools and methods) are combining artificial intelligence (AI) knowledge representations
and programming techniques with the intent to deliver instructional content and support tailored to the individual
(Conati & Manske, 2009). From a holistic perspective, such tools and methods personalize training by considering
an individual‟s historical data, real-time behavior, and cognitive measures to predicting comprehension levels and
affective states (i.e. frustration, boredom, excitement). This historical and real-time interpretation of the trainee is
used for concurrent adaptation of pedagogical and feedback strategies within training content.
Several ITS studies within academic settings report significant learning gains among students receiving adaptive ITS
support when compared to students in a traditional schoolhouse environment (Koedinger, Anderson, Hadley &
Mark, 1997; Kulik & Kulik, 1991). However, the majority of those systems supported domains with well-defined
problems that require well-defined solutions (i.e. physics, algebra). With recent trends in virtual scenario-based
training in the defense and medical communities, there has been a major push to simulate more ill-defined tasks that
require critical decision-making and swift problem-solving. Primary issues associated with ill-defined scenario
training are the lack of a suitable design framework, and determining an appropriate level of support/direction
through pedagogy and feedback. This paper will compare ITS pedagogical design considerations between well-
defined and ill-defined tasks, identify the variables of interest that have the greatest impact on performance and skill
acquisition, and present a high-level design architecture.
ABOUT THE AUTHORS
Mr. Benjamin S. Goldberg is a member of the Learning in Intelligent Tutoring Environments (LITE) Lab at the
U.S. Army Research Laboratory's (ARL) Simulation and Training Technology Center (STTC) in Orlando, FL. He
has been conducting research in the Modeling and Simulation community for the past 3 years with a focus on
adaptive learning and how to leverage Artificial Intelligence tools and methods for adaptive computer-based
instruction. Mr. Goldberg is a Ph.D. student at the University of Central Florida and holds an M.S. in Modeling &
Simulation. Prior to employment with ARL, he held a Graduate Research Assistant position for two years in the
Applied Cognition and Training in Immersive Virtual Environments (ACTIVE) Lab at the Institute for Simulation
and Training. Recent publications include a proceedings paper in the 2010 International Conference on Applied
Human Factors and Ergonomics and a poster paper presented at the 2010 Spring Simulation Multiconference.
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 2 of 12
Dr. Heather Holden is a researcher in the Learning in Intelligent Tutoring Environments (LITE) Lab within the
U.S. Army Research Laboratory – Human Research & Engineering Directorate (ARL-HRED). The focus of her
research is in artificial intelligence and its application to education and training; technology acceptance and Human-
Computer Interaction. Dr. Holden‟s doctoral research evaluated the relationship between teachers' technology
acceptance and usage behaviors to better understand the perceived usability and utilization of job-related
technologies. Her work has been published in the Journal of Research on Technology in Education, the International
Journal of Mobile Learning and Organization, the Interactive Technology and Smart Education Journal, and several
relevant conference proceedings. Her PhD and MS were earned in Information Systems from the University of
Maryland, Baltimore County. Dr. Holden also possesses a BS in Computer Science from the University of Maryland,
Eastern Shore.
Mr. Keith Brawner is a researcher for the Learning in Intelligent Tutoring Environments (LITE) Lab within the U.
S. Army Research Laboratory - Human Research & Engineering Directorate (ARL-HRED). He has 4 years of
experience within U.S. Army and Navy acquisition, development, and research agencies. He holds a Masters degree
in Computer Engineering with a focus on Intelligent Systems and Machine Learning from the University of Central
Florida, and is currently in pursuit of a doctoral degree in the same field. The focus of his current research is in
machine learning, adaptive training, student modeling, and knowledge transfer.
Dr. Robert Sottilare is the Associate Director for Science & Technology within the U.S. Army Research
Laboratory - Human Research & Engineering Directorate (ARL-HRED). He has over 25 years of experience as both
a U.S. Army and Navy training & simulation researcher, engineer and program manager. He leads STTC's
international program and participates in training technology panels within both The Technical Cooperation Program
(TTCP) and NATO. He has a patent for a high resolution, head mounted projection display (U.S. Patent 7,525,735)
and his recent publications have appeared in the Journal for Defense Modeling and Simulation, the NATO Human
Factors and Medicine Panel's workshop on Human Dimensions in Embedded Virtual Simulation and the Intelligent
Tutoring Systems Conference. Dr. Sottilare is a graduate of the Advanced Program Managers Course at the Defense
Systems Management College at Ft. Belvoir, Virginia and his doctorate in modeling & simulation is from the
University of Central Florida. The focus of his current research program is in machine learning, trainee modeling and
the application of artificial intelligence tools and methods to adaptive training environments.
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 3 of 12
Enhancing Performance through Pedagogy and Feedback: Domain
Considerations for ITSs
Benjamin S. Goldberg, Heather K. Holden, Keith Brawner, and Robert A. Sottilare
U.S. Army Research Laboratory – Human Research and Engineering Directorate
Orlando, Florida
{benjamin.s.goldberg; robert.sottilare; heather.k.holden; keith.w.brawner}@us.army.mil
INTRODUCTION
Pedagogy and instructional design are fundamental
components to successful training implementation. This
is true more than ever with the incorporation of
computer-based training platforms that promotes and
mediates self-directed learning. Pedagogy focuses on
adaptive instruction that individualizes training
experiences to the needs of a particular learner. Based
on knowledge, competency, and state, training is
tailored to meet skill level and feedback/interventions
are incorporated as tutoring mechanisms to aid in
restoring or maintaining a positive learning state. Such
methods are being pursued by the military and medical
communities to instantiate alternative solutions to
expensive and resource straining live exercises. To gain
the full benefits of such techniques, design
considerations need to incorporate qualities of
instructor characteristics that provide real-time
performance assessment and feedback. Intelligent
Tutoring Systems (ITS) are one such approach that use
Artificial Intelligence (AI) knowledge representations
with machine learning techniques for the purpose of
producing concurrent performance and state (cognition
and affect) diagnoses on the individual and team level
(Conati & Manske, 2009). However, an issue with this
approach is knowing when and how to adapt training
content when an individual is classified in a negative
“readiness to learn” state.
Much of the research conducted to answer this question
involves studying the tactics performed by human
tutors. Tutors are found to be most effective because
humans are able to adapt and individualize instruction
based on the particular needs of a trainee (Lane &
Johnson, 2008). This involves an active balance of
participation by the trainee with guidance facilitated by
the tutor. The driving factors for determining guidance
are based on competency exhibited through
performance and dynamic state variables (e.g.,
engagement, frustration, boredom, confusion, etc.) that
fluctuate during interaction. The objective is to have a
trainee perform as much of a task as possible while a
tutor provides constructive feedback aimed to minimize
frustration and confusion (Merrill, Reiser, Ranney, &
Trafton, 1992).
Though there has been empirical evidence of improved
performance among ITSs, the results still do not meet
or exceed the effectiveness of human tutors. This is due
in part to a human‟s ability to read and interpret cues
linked with affective states associated to cognitive
performance. For selection of the most advantageous
pedagogical strategies, the system must know how the
interacting trainee is feeling cognitively and
emotionally to adapt content that matches their current
state. To produce computer-based platforms that
exhibit the same benefits seen in one-to-one human
instruction (Bloom, 1984), the system must be able to
make state determinations as well as or better than that
of a person. A large amount of experimentation has
been conducted incorporating sensor technology that
monitors both behavioral and physiological markers
believed to be correlated with cognitive and affective
states to make this a realization (D‟Mello, Taylor, &
Graesser, 2007; Berka et al, 2007; McQuiggan, Lee, &
Lester, 2007; Ahlstrom & Friedman-Bern, 2006). From
a holistic perspective, such tools and methods
personalize training by considering an individual‟s
historical data, real-time behavior, and cognitive
measures to predicting comprehension levels and
affective states. This historical and real-time
interpretation of the trainee is used for synchronized
adaptation of pedagogical and feedback strategies
within training content to maintain appropriate
challenge and to curb boredom.
Yet, decisions on how to adapt content based on state
assessment follows few standards. With hundreds of
theories on instructional design (see Marzano, 1998;
Marzono, 2003; Bransford, Brown, & Cocking, 1999),
there are no deemed best practices for adaptive
instruction. This paper will highlight design
considerations based on domain characteristics and will
present a high-level pedagogical architecture. The
architecture is based on empirical evidence from past
studies in the field and theoretical perspectives on
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 4 of 12
instructional design geared for computer-based
education. Appropriate instructional strategy selection
requires analysis of a number of variables that drive
task mechanics. Establishing a defined framework that
assists in decomposing task components for
instructional design can improve adaptive capabilities
and reduce time in transitioning systems to training
houses.
Trends of ITS Implementation
ITS research aims to make training environments
adaptable to different learning needs and abilities on an
individual level. Majority of systems control the user
experience through interacting models that correspond
with the elements utilized by a human tutor. This
includes knowledge about the student (Student Model),
instructional strategy selection rules (Pedagogical
Model), knowledge about the domain being trained
(Domain Model), and knowledge of how to
successfully perform domain tasks (Expert Model)
(Durlach & Ray, 2011). The type of data fed into these
models is dependent on the domain and available
sensing technology designed into the platform.
Current fielded applications in academia using ITS
technologies have shown significant learning gains over
long-established instructional methods, with the best
platforms producing an average increase of 1.0
standard deviation over conventional practices
(Anderson, Corbett, Koedinger, & Pelletier, 1995;
VanLehn, Lynch, Schulze, Shapiro, Taylor, & Treacy,
2005; Koedinger, Anderson, Hadley & Mark, 1997;
Kulik, 2003). However, these successful ITS
applications are administered within well-defined
domains (e.g., math, physics, chemistry), which involve
specific procedures for satisfying task objectives. In
this context, performance is easily assessable based on
models of expert performance. When actions
performed delineate from successful routines, feedback
and/or content manipulations (change of pace and/or
difficulty) are administered to reduce errors and
enhance training transfer. Other approaches for well-
formed domains that apply artificial intelligence and
cognitive science for adaptation include production
systems, case-based reasoning, Bayesian networks,
theorem proving, and constraint satisfaction algorithms
(Shaffer & Graesser, 2010).
Because performance alone cannot accurately gauge
overall training effectiveness, new directions are being
taken to enhance the capabilities of ITS components.
Incorporating mechanisms that can track affective
states among individual trainees will improve the
diagnostic capacity of such systems to classify
emotional responses that may hinder learning (i.e.,
boredom, frustration, fatigue). Research efforts are
looking at better ways to collect data on the trainee for
accurate real-time assessment that can be used to tailor
training to match strengths and weaknesses. This
includes indentifying techniques that can passively
gather information while remaining unobtrusive to the
individual. With new data streams being fed into the
learner model, pedagogical decision functions can
leverage this information to understand more about the
current state of the trainee. If such a platform can
diagnose competence/performance, motivation, and
emotional response, intervention selection can facilitate
multiple options outside of feedback driven by
performance.
FRAMEWORK FOR DOMAIN DEFINITION
There are a number of variables to consider when
adapting a training experience, with domain explicates
being a major influencing factor. Dependent on the
process and complexity, instructional and pedagogical
design practitioners need to decompose task actions
into basic functional components. This is led by defined
objectives training aims to prepare. The process to
achieve task objectives is based on the structure of
actions required for successful performance, as well as
how well the initial state and goal state are specified
(Goel, 1995). If the structure of actions follow specific
standards and involves the same process in each
instantiation, the task components are considered well-
defined. In the instance where procedures are
ambiguously defined and there are no clear set actions
for meeting goal objectives, the domain of interest is
considered ill-defined. This is further described as
problems that have less specific criteria for determining
when an objective has been satisfied and all
information required for a solution is not supplied
(Simon, 1973). The domain definition must also take
into account the complexity of the task being
performed. Complexity is comprised of how difficult
(easy/hard) the task is to conduct, the type of
environment it‟s performed within, and the extraneous
factors (weather, opposition, visibility, etc.) that may
influence its outcome.
Well-defined domains that score high in complexity
often times require training of skills outside specific
task execution. In the context of military readiness
training, exhibiting performance standards is essential.
However, independent of their military occupational
specialties (MOS), Soldiers are required to demonstrate
higher order thinking skills that exhibit the capability of
adapting decision-making tactics in unstable
environments where situations and conditions rapidly
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 5 of 12
change (US Army Training and Doctrine Command,
2000). The United States Military Academy‟s (USMA)
Center for Enhanced Performance identified the
following elements as critical for performance
improvement among warfighters: metacognitive
awareness, attentional control, goal-setting, stress
management, and visualization (Zinsser, Perkins,
Gervais, & Burbelo, 2004). Zinsser et al (2004) further
state that establishing these competencies empower
individual Soldiers to create efficient thinking habits;
improve attentional resources for enhanced situational
awareness; and provides experience for coping with
physical, emotional, and mental responses during high-
demanding tasks. Though many MOS tasks may follow
well-defined routines, extraneous factors can
significantly impact the procedure or environment the
task is being performed within. Because of this,
personnel must be able to adapt in real-time to ensure
objectives are reached. Training these competencies
among all Soldiers is vital for an adaptive Force.
Instilling these elements in trainees requires effective
and efficient training paradigms. Providing training that
achieves efficient acquisition of these elements is not
easily administered. Defining performance criteria for
such skills is difficult and is based on specific scenario
interactions. Because of this ill-defined classification,
standards need to be developed that highlight key
strategies and adaptation approaches for design
practitioners to enhance system support features for
maximizing training effectiveness. With a push by the
military for a learner-centered approach to training,
tools and methods that promote self-directed learning
are required (TRADOC, 2011). TRADOC further
identifies the need for pursuing adaptive training and
tutoring technologies and to develop standards for
implementing these capabilities in computer-based
platforms. Yet, a number of issues must be addressed in
ITS design to facilitate development of the critical
competencies associated with desired training
outcomes.
With a two dimensional approach to domain definition,
instructional strategies can be specified based on the
component characteristics identified within the domain
designation. Through task analysis, procedure and
complexity can be classified and used for formulation
of training objectives. Based on competency and
experience, specific training objectives are tailored on
the individual level to promote efficient progression
from novice to expert. As a trainee progresses through
initial content, training objectives are adapted to
introduce extraneous factors that will influence
execution. Time spent on training, progression tactics
from basic functional training to skill mastery,
repetition and remediation techniques, and technology
aids all play an important role in training
implementation (Zipperer, Klein, Fitzgeral, Kinnison &
Graham, 2003).
DESIGN CONSIDERATIONS DEPENDENT OF
DOMAIN
Considerations for pedagogy and feedback are founded
on empirical evidence from past studies using
computer-based learning environments and classic
learning theory literature. Instructional strategy
selection and feedback implementation are the
variables of interest, with a goal to discern those that
have the highest impact on learning outcomes.
Distinguishing a list of best practices based on domain
definition is premature at this point and requires
rigorous empirical evaluations of the following
findings, testing their validity across multiple domains.
This paper aims to identify the current state of
pedagogy and feedback research in the ITS community
and to identify the strategies that have shown
significant improvements over traditional instruction or
computer-based systems that lack adaptive faculties.
The remainder of the section will review considerations
instructional designers must reflect on when developing
an ITS application. This requires sound design
procedures that takes into account all elements
associated with a given training event and the interface
components used. As mentioned above, learning
objectives are strongly tied to domain definition. Based
on the categorized domain a given training objective
falls within, methods for pedagogical design need to be
identified based on the type of actions being performed.
Following domain definition based on task procedure
and complexity, an instructional designer should follow
a routine process to create the training experience. Five
rudiments have been identified that must be addressed
in adaptive training design:
Curriculum: Specifies explicit content for
training. This involves scenario design
through task designation and decomposition.
Instructional Strategy: Highlights how content
will be presented, the pace of training, the
types of actions and procedures performed by
the trainee, and scenario difficulty/complexity.
Performance Measures: A challenge
associated with training is defining what is
deemed as successful performance. Variables
must be identified that measure performance
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 6 of 12
and are congruent with positive training
transfer for the specific outcome of interest.
Pedagogical Interventions: Determines
feedback and support considerations
associated to the learning objectives and
„readiness to learn‟ state. This includes
manipulations of content and feedback
interventions.
Student Model Data: Highlights specific data
that is needed for cueing interventions.
Decision functions must also be defined that
trigger adaptive interventions (performance
data vs. cognitive/affective state
determinants). Data fed into model will vary
depending on system functionalities.
Each aforementioned element requires awareness
during the design phase of an ITS application. The
following subsections will review empirical studies of
ITS applications within both well-defined and ill-
defined domains. The review will highlight
instructional and feedback implementation strategies
and their impact on learning outcomes when compared
to control settings. This effort aims to identify
similarities among successful empirically tested
systems and to categorize a domain definition with
sound techniques for administering adaptive training.
Well-Defined Domains
Instructional strategies of existing ITSs are based on
human teaching, informed by learning theories, and are
facilitated by technology (Woolf, 2009). Four strategies
based on human instruction commonly used in ITSs are
apprenticeship training, problem solving, tutorial
dialogue, and collaborative learning (Woolf, 2009).
Tutors that cater to well-defined domains can
incorporate any combination of these strategies. ITSs
that use apprenticeship training contain an expert model
to track student performance, provide advice on
demand, and support multiple avenues to solutions.
Such tutors will scaffold instruction as needed, but will
often stay in the background having the student be
more responsible for their performance. This strategy is
ideal for tasks that are best learned by doing. Sherlock
is an apprenticeship training-based environment that
simulated the structure/function of a complex electronic
diagnostic board (Lesgold et al., 1992). Although
Sherlock only provided feedback when requested by
the learner, the scores of those who used the
environment increased approximately 35% when
compared to learners who received no feedback
(Corbett, Koedinger, & Anderson, 2007).
Problem-solving is another traditional ITS instructional
strategy that uses error-handling techniques and
production rules to navigate instruction. For example,
the Andes physics tutor uses problem definition,
physics rules, a solution graph, action interpreter, and a
help system to provide procedural, conceptual, and
example guidance (Gertner & VanLehn, 2000). This
tutor has been shown to increase scores by one standard
deviation (Schulze, Shelby, Treacy, Wintersgill,
VanLehn, & Gertner, 2000) and one-third of a letter
grade (Gertner & VanLehn, 2000). Like most problem-
solving tutors for well-defined domains, the Andes
tutor uses model-tracing techniques to monitor the
students‟ progress through a problem solution. In
model-tracing, the tutor tries to infer the process by
which a student arrived to a solution and uses that
inference as the basis for remediation. Model-tracing
tutors typically contain expert (production) rules, buggy
rules (for error handling), a model tracer, and a user
interface (Kodagnallur, Weitz, & Rosenthal, 2005).
Although model-tracing tutors have shown to increase
student performance, their adaptation and
accountability for individual differences is significantly
limited. Traditional model-tracing tutors do not allow
for new questions or multi-step lines of questioning.
However, second/third generation model-tracing tutors
are created to better personalize instruction. For
example, the Pump Algebra Tutor (PAT) is a problem-
solving tutor that includes a cognitive (psychological)
model to assess the process of cognition behind
successful and near-successful student performance.
PAT also uses knowledge tracing to monitor student‟s
learning from problem to problem. This technique
identifies students‟ strengths and weaknesses relative to
the cognitive model‟s production rules (Koedinger,
Anderson, Hadley, & Mark, 1997). PAT is used within
thousands of schools and has found to improve student
performance on standardized test by 15-25 %
(Koedinger & Corbett, 2006). Ms. Lindquist, another
model-tracing based tutor for algebra, added a „tutorial
model‟ that provides the capability asking questions
and thinking about the knowledge behind the next
problem solving step. Ms. Lindquist allows for the
ability of engaging in dialog with the learner
(Heffernan, Koedinger, & Razzaq, 2008).
Well-defined model-tracing tutors previously
mentioned have improved their adaptability to learner
cognition and have been shown to increase student
performance; however, they still do not cater to natural
interactions seen in one-to-one tutoring. Other ITSs,
such as AutoTutor (Graesser, Chipman, Haynes, &
Olney, 2005), incorporate natural language interfaces
that allows spoken dialogue and adaptation. This
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 7 of 12
promotes collaborative inquiry learning, which has
been shown to increase student performance and other
learner outcomes.
Ill-Defined Domains
To make ITSs effective across a number of domains,
focused feedback and scenario adaptations are required
that assist trainees in knowledge/skill acquisition.
Empirical studies have been conducted looking at
varying pedagogical approaches and to view their effect
on learning outcomes. The issue with providing real-
time feedback is identifying the mechanisms and
decision functions that trigger an intervention, and
designing environments that guide trainees to optimal
interactions without limiting performance. This requires
user models that account for the uncertainty,
dynamicity, and multiple interpretations of how to
execute ill-defined tasks (Lynch, Ashley, Mitrovic,
Dimitrova, Pinkwart, & Aleven, 2010).
Pedagogy within ill-defined domains takes many forms,
each of which applies theoretical underpinnings
believed to maximize training effectiveness. The
underlying issue is definitive feedback often given for
well-structured domains are difficult to provide in an
ill-defined setting (Walker, Ogan, Aleven, & Jones,
2008). Techniques such as model-tracing, expert
systems, and constraint-based reasoning are not optimal
in this context because they lack the specifications to
support tutoring services outside of task performance
(Bratt, 2009). Expert systems have the potential to be
leveraged for such training, but comparing a trainee‟s
solution against an ideal solution does not always
provide an explanation or reasoning of how the result
was constructed (Fournier-Viger, Nkambou, Nguifo, &
Mayers, 2010). Depending on the nature of the task and
the learning objectives training aims to prepare,
specialized instructional delivery and feedback
mechanisms must be designed to facilitate training in
ill-specified problem spaces.
The challenge is defining a solution path that caters to
learning critical elements associated with conducting
ill-defined tasks. Lynch, Ashley, Pinkwart, and Aleven
(2008) proposes solution paths for ill-defined domains
are constructed through: (1) multiple characterizations
of a problem to specify components and constraints that
have been undefined for selection and discrimination
among alternatives, (2) experience for adapting to
second and third order effects given the context of a
specific problem space, and (3) to justify scenario
actions taken with concepts and principles linked to
training curriculum. The intent of this approach is to
develop a trainee‟s knowledge and reasoning skills so
to avoid the worst outcomes when choosing an action
response (Bratt, 2009). Development of such skills is
through practice, and simulation-based training
environments offer a low-cost alternative to running
live exercises that are often resource straining.
Simulated training experiences promote the
constructivist and experiential methodologies of
learning by facilitating a trainee to solve multiple
problems of varying complexity across a number of
situations (Bratt, 2009; Raybourn, 2007). However, this
approach requires significant instructional guidance at
times to avoid negative transfer. To enhance the
capabilities of simulation-based platforms used solely
as practice environments, adaptive intelligent tutors
have been added in systems to aid in instilling the
critical aspects of performance, to ensure trainees avoid
practicing mistakes or displaying misconceptions, and
to reduce the role of the instructor (Thomas & Milligan,
2004). A workshop held at the 9th
International
Conference on Intelligent Tutoring Systems identified
the following explicit domains as ill-defined, which
require specialized feedback considerations: medical
diagnosis and treatment, intercultural relations and
negotiations, inquiry learning, ethical reasoning,
robotics operation, and object-oriented design (Lynch
et al, 2008). Each of these domains requires higher
order thinking skills that enable an individual to
perform decision-making, problem solving and goal-
conflict resolution.
Five tutorial strategies commonly used for development
of higher order thinking skills are question prompts,
clarification, hints, examples, and redirection (Alvarez-
Xochihua, Bettati, & Cifuentes, 2010). Each strategy is
intended to support problem-solving situations and
promote metacognitive awareness. An ITS designed for
training problem solving with cybersecurity personnel
incorporated a Mixed-Initiative framework for
providing feedback (Alvarez-Xochihua et al, 2010).
The framework was applied to a case-based
instructional system that allowed execution of problem
solving techniques across a number of distinct
scenarios. The system was designed as a re-active
platform that responds to student requests. Based on
state-determinations generalized from interactions prior
to the help request, the mixed-imitative component of
the ITS decides the format of feedback to deliver from
the five strategies listed above. The system is currently
being empirically evaluated to assess if the mixed-
initiative approach provides relevant feedback.
Another approach to feedback implementation in ill-
defined domains is providing adaptive guidance during
scenario interaction and during an after-action review
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 8 of 12
(AAR) for the purpose of promoting reflection. This
approach was implemented in ELECT BiLat, a game-
based trainer used for teaching and practicing cultural
awareness and negotiating skills (Lane, Core, Gomboc,
Karnavat, & Rosenberg, 2007). The environment is
designed around interactive narrative between the
trainee and artificial agents that respond to the
exchanges taken by the system user. Feedback is
determined by an expert model. Based on the current
state of a scenario, the expert model runs a search
algorithm that identifies all available action selections,
filters out actions not appropriate for meeting
objectives, filters out actions previously performed, and
identifies the action selections congruent with expert
performance (Lane et al, 2007). Based on the trainee
dialogue selection, feedback is generated providing
either a hint, a positive remark for a good action, or
negative feedback with a short explanation (Lane,
Hays, Core, Gomboc, Forbell, Auerbach, & Rosenberg,
2008). Study results comparing effectiveness of
feedback implementation in comparison to no coaching
showed 89% participants who received real-time
guidance completed the scenario successfully while
only 59% who received no feedback met training
objectives (Lane et al, 2008).
Natural dialogue-based tutoring is an additional method
for providing real-time feedback in an ITS
environment. Systems designed for well-defined
domains have shown success incorporating natural
language dialogue that tasks a trainee with articulating
the reasoning process during scenario execution.
AutoTutor, an ITS that teaches introductory computer
literacy, showed an increase in performance of 0.5
standard deviations in comparison to learners who
received instruction from a text book (Graesser,
Wiemer-Hastings, Wiemer-Hstings, Kreuz, & Tutoring
Research Group, 1999). The system prompts trainees to
provide explanations for „How‟, „Why‟, and „What-if‟
questions. A concern with this approach is the accuracy
of dialogue systems and their effect on training
performance. A study was conducted with AutoTutor to
gage the effect speech recognition errors had on
learning, with the results conveying a subtle impact on
performance as well as on a participant‟s emotions and
attitudes (D‟Mello, King, Stolarski, Chipman &
Graesser, 2007).
Natural language dialogue is now being integrated in
ITSs geared for ill-defined domains. The EER-Tutor
was designed to instruct individuals on database design
by utilizing a hands on practice environment that
incorporates dialogue interaction with a computer-
based tutor (Weerasinghe, Mitrovic & Martin, 2009).
Tutor interventions are applied when an error is present
and is facilitated by tutorial dialogues. An independent
dialogue was designed for each identified error type
and feedback was implemented by a rule-based
reasoning system. Interventions were designed to
facilitate remediation, aid in completing the session and
assisting with technical problems, and for helping with
the interface components (Weerasinghe et al, 2009).
Feedback is also designed to be adaptive based on a
learner‟s current domain knowledge and reasoning
skills. The model proposed for this tutor was evaluated
by five acting judges to rate the appropriateness of
feedback selection and timing. Conclusions from this
study supported the models ability for error-
remediation in ill-defined tasks and indicated that
trainees acquired domain concepts in the natural
dialogue sessions.
Because of the uncertainty in performing ill-defined
tasks, knowing the appropriate pedagogy to enact is
difficult to determine. As can be seen in this paper, the
ITS research community is taking multiple avenues to
facilitate adaptive instruction across well- and ill-
defined domains. Building standards and guidelines for
pedagogical intervention selection based on defined
training objectives can enhance systems to be
interoperable across a number of domains and aid in
reducing time on instructional design. This requires a
modular architecture that incorporates all ITS
components for the purpose of guiding feedback
selection.
ITS PEDAGOGY ARCHITECTURE BASED ON
DOMAIN CONSIDERATIONS
The Generalized Intelligent Framework for Tutoring
(GIFT) architecture (see Figure 1) introduces the
components of sensor, trainee, pedagogical, learning
management system (LMS), and domain modules. A
domain module in GIFT processes feedback requests
and scenario changes by adjusting scenario elements or
user interface components. It allows for the pass-
through of domain independent interventions, and
allows the domain to respond to specific hint requests.
A domain module can assess feedback either through a
priori knowledge of the correct answer, having the
ability to calculate the correct answer, or through
comparison of a built-in expert model, depending on
how well-defined a domain is. The architecture
supports a hybrid model approach to feedback selection
and requires production rules that dictate activation.
Through system use the production rules will be
iteratively updated based on the data fed into the
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 9 of 12
Figure 1: Generalized Intelligent Framework for Tutoring (GIFT)
student and domain models. Variables of performance,
competency, affect, and cognition will all be
determinants of implementing a training intervention.
The architecture will allow for grouping of
feedback/manipulation strategies with a triggering
variable and supports empirical evaluations to test their
effect on training outcomes.
The primary outputs of pedagogical strategy decisions
are whether to make an intervention, and the next
instructional content which must be presented. While
the instructional content decisions are processed out of
view of the user, interventions have a sizable effect.
Interventions take one of two forms, either that of a
domain-specific feedback such as “aim higher”,
domain-independent feedback such as an emotional, or
a metacognitive prompt. Decisions to change
instructional content also come in different forms.
Content decisions can modify task demand, modify task
complexity, or change the types of content presented. A
well-designed domain-specific component must address
these things.
In order for pedagogy and feedback to be successful,
the architecture requires a domain module that supports
the types of feedback requested. While the vast
majority of the components of an ITS may be made
domain independent, there must always be a specific
component of the architecture to deal with the problems
that the instructor desires to teach. The fundamental
problems of domain dependent components are how to
assess student actions, how to respond to instructional
changes, how to respond to requests for immediate
feedback, and an interface which supports learning
(Sottilare, Holden, Goldberg, & Brawner, In Review).
The architecture designed must have built-in support
for these types of instructional activities.
In a GIFT prototype system for the well-defined
domain of addition, this module is being constructed in
the following manner. Student action assessment is
based on whether or not each digit in a multi-digit
number is computed correctly. Feedback generation is
handled either through a pass-through emotional
prompt or buggy mistake prompt. Complexity is
handled by adding or subtracting digits to the numbers
to be added, while task demand is changed by allowing
less time for the problem to be solved. The user
interface is a simple screen with the ability to input
added numbers. This is a proof-of-concept system that
shows the ease of use of inputting a given domain into a
more generalized architecture.
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 10 of 12
CONCLUSIONS AND FUTURE WORK
While we intuitively know that it is better to have more
information when we are making decisions to tailor
instructional feedback and content to individual trainee
needs, the influence of specific trainee attributes on
instructional decisions can be debated. Additional
experimentation is needed to quantify the impact of
trainee attributes. For example, the importance of
personality attributes like openness to performance
might differ by task type (e.g., ill or well-defined tasks;
individual or collective tasks).
Additionally, implementing sound pedagogy in
computer-based training will require accurate state
classifications that will determine timing and type.
Research is needed to evaluate the influence of macro
and micro variables in classifying trainee cognition and
affect. Macro variables are generally known at training
exercise start and include trainee states like affect (e.g.,
personality), domain competence, learning preferences
and demographic data (e.g., gender, training history).
Micro variables include real-time behavioral and
physiological attributes. Behavioral data may be
captured by recording interactions within the training
simulation (e.g., mouse movement or control selection)
or through sensor methods (e.g., motion capture).
Physiological attributes (e.g., interbeat heartrates,
brainwaves) are generally captured via sensor methods.
It will be critical to build validated models of trainee
cognitive and affective states using behavioral and
physiological measures, but to make the use of these
models practical for military training, it will be
essential to develop sensor methods that are: low-cost,
unobtrusive and portable. Empirical evaluations,
validating classification models, and feedback
approaches across domains is required for developing
standards of feedback implementation in adaptive
training. This will enable congruent approaches across
domains in terms of pedagogical approaches based on
domain definition and state assessment.
REFERENCES
Ahlstrom, U., & Friedman-Bern, F.J. (2006). Using eye
movement activity as a correlate of cognitive
workload. International Journal of Industrial
Ergonomics, 36(7), 623-636.
Alvarez-Xochihua, O., Bettati, R. & Cifuentes, L.
(2010). Mixed-initiative intelligent tutoring
addressing case-based problem solving. Technical
Report TAMU-CS-TR-2010-7-2, Texas A&M
University.
Anderson, J. A., Corbett, A. T., Koedinger, K., &
Pelletier, R. (1995). Cognitive Tutors: Lessons
Learned. The Journal of the Learning Sciences, 4(2),
167-207.
Berka, C., Levendowski, D.J., Lumicao, M.N., Yau, A.,
Davis, G., Zivkovic, V.T., Olmstead, R.E.,
Tremoulet, P.D., Craven, P.L. (2007). EEG
correlates of task engagement and mental workload
in vigilance, learning, and memory tasks. Aviation
Space and Environmental Medicine; 78(5,
Suppl.):B231-B244.
Bloom, B. S. (1984). The 2 sigma problem: The search
for methods of group instruction as effective as one-
to-one tutoring. Educational Researcher, 13(6), 4-
16.
Bransford, J. D., Brown, A. L., & Cocking, R. R. (Eds.).
(2000). How people learn: Brain, mind, experience,
and school. Washington, DC: National Academy
Press.
Bratt, E.O. (2009). Intelligent tutoring for ill-defined
domains in military simulation-based training.
International Journal of Artificial Intelligence in
Education, 19, 337-356.
Conati, C., & Manske, M. (2009). Evaluating Adaptive
Feedback in an Educational Computer Game. In Z.
Ruttkay, M. Kipp, A. Nijholt, & H.H. Vilhjálmsson
(Eds.), Proceedings of 9th International Conference
on Intelligent Virtual Agents (IVA ’09), LNCS (Vol.
5773, pp. 146-158). Berlin: Springer.
Corbett, A. T., Koedinger, K. R., & Anderson, J. R.
(1997). Intelligent tutoring systems. In M. G.
Helander, T. K. Landauer, & P. Prabhu (Eds.),
Handbook of Human-Computer Interaction, 2nd
edition (Chapter 37). Amsterdam, The Netherlands:
Elsevier Science.
D‟Mello, S. K., Taylor, R., & Graesser, A. C. (2007).
Monitoring Affective Trajectories during Complex
Learning. In D. S. McNamara & J. G. Trafton (Eds.),
In Proceedings of the 29th Annual Cognitive Science
Society (pp. 203-208). Austin, TX: Cognitive
Science Society.
D‟Mello, S.K., King, B., Stolarski, M., Chipman, P. &
Graesser, A. (2007). The effects of speech
recognition errors on learner‟s contributions,
knowledge, emotions, and interaction experience. In
Proceedings of the SLaTE Workshop on Speech and
Language Technology in Education. Farmington,
PA, pp. 49-52.
Fournier-Viger, P., Faghihi, U., Nkambou, R., &
Nguifo, E. (2010). ITS in Ill-Defined Domains:
Toward Hybrid Approaches. In V. Aleven, J. Kay,
& J. Mostow (Eds.), Proceedings of the 10th
International Conference on Intelligent Tutoring
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 11 of 12
Systems (ITS ’10), LNCS (Vol. 6095, pp. 318-320).
Berlin: Springer.
Gertner, A., & VanLehn, K. (2000). Andes: A coached
problem solving environment for physics. In G.
Gauthier, C. Frasson, & K. VanLen (Eds.),
Proceedings of the 5th
International Conference on
Intelligent Tutoring Systems (ITS ’00), LNCS (Vol.
1839, pp. 133-142). Berlin: Springer.
Goel, V. (1995). Sketches of Thought. Cambridge, MA:
MIT Press.
Graesser, A., Wiemer-Hastings, K., Wiemer-Hastings,
P., Kreuz, R., & the Tutoring Research Group
(1999). AUTOTUTOR: A simulation of a human
tutor. Journal of Cognitive Systems Research, 1(1),
35-51.
Graesser, A., Chipman, P., Haynes, B., and Olney, A.
(2005). AutoTutor: An intelligent tutoring system
with mixed-initiative dialogue. IEEE Transactions on
Education, 48(4), pp. 612-618.
Heffernan, N., Koedinger, K., & Razzaq, L. (2008).
Expanding the model-tracing architecture: A 3rd
generation intelligent tutoring tutor for Algebra
symbolization. The International Journal of Artificial
Intelligence in Education, 18(2), 153 – 178.
Kodaganallur, V., Weitz, R., & Rosenthal, D. (2005). A
comparison of model-tracing and constraint-based
intelligent tutoring paradigms. International Journal
of Artificial Intelligence in Education, 6(2), 117-144.
Koedinger, K. R., Anderson, J. R., Hadley, W. H., &
Mark, M. (1997). Intelligent tutoring goes to school
in the big city. International Journal of Artificial
Intelligence in Education, 8, 30-43.
Koedinger, K.R., & Corbett, A. T. (2006). Cognitive
Tutors: Technology bringing learning science to the
classroom. In Sayer, K. (Ed.). The Cambridge
Handbook of the Learning Sciences (pp. 61-78).
Cambridge, MA: Cambridge University Press.
Kulik, J.A. (2003). Effects of using instructional
technology in elementary and secondary schools:
What controlled evaluation studies say. Final Report
(SRI Project No. P10446.001). Arlington, VA: SRI
International
Lane, H. (2006). Intelligent tutoring systems: Prospects
for guided practice and efficient learning. In Army‟s
science of learning workshop. Hampton, VA.
Lane, H.C., Core, M.G., Gomboc, D., Karnavat, A. &
Rosenberg, M. (2007). Intelligent tutoring for
interpersonal and intercultural skills. In Proceedings
of the Interservice/Industry Training, Simulation,
and Education Conference (I/ITSEC 2007), Orlando,
FL.
Lane, H.C., Hays, M., Core, M., Gomboc, D., Forbell,
E., Auerbach, D., & Rosenberg, M. (2008). Coaching
intercultural communication in a serious game. In
T.W. Chan (Ed.) Proceedings of the 16th
International Conference on Computers in
Education (pp. 35-42). Jhongli City, Taiwan:
APSCE.
Lane, H.C., & Johnson, W.L. (2008). Intelligent
Tutoring and Pedagogical Experience Manipulation
in Virtual Learning Environments. In J. Cohn, D.
Nicholson, & D. Schmorrow (Eds.), The PSI
Handbook of Virtual Environments for Training and
Education (pp. 393-406). Westport, CT: Praeger
Security International.
Lesgold, A., Lajoie, S., Bunzo, M., & Eggan, G. (1992).
SHERLOCK: A coached practice environment for an
electronics troubleshooting job. In J. Larken & R.
Chabay (Eds.), Computer-Assisted Instruction and
Intelligent Tutoring System: Shared Issues and
Complementary Approaches (pp. 201-238).
Hillsdale, NJ: Lawrence Erlbaum.
Lynch, C., Ashley, K.D., Pinkwart, N., & Aleven, V.
(2009). Concepts, structures, and goals: redefining
ill-definedness. International Journal of Artificial
Intelligence in Education, 19, 253-266.
Lynch, C., Ashley, K., Mitrovic, T., Dimitrova, V.,
Pinkwart, N., & Aleven, V. (2010). Intelligent
Tutoring Technologies for Ill-Defined Problems and
Ill-Defined Domains. In Proceedings of the 4th
International Workshop on Intelligent Tutoring
Systems and Ill-Defined Domains at the 10th
International Conference on Intelligent Tutoring
Systems (ITS2010), Pittsburgh, PA.
Marzano, R. J. (1998). A theory-based meta-analysis of
research on instruction. Aurora, CO: Mid-continent
Research for Education and Learning. (Eric
Document Reproduction Service No. ED 427 087)
Marzano, R. J. (2003). What works in schools:
Translating research into action. Alexandria, VA:
Association for Supervision and Curriculum
Development.
McQuiggan, S., Lee, S., & Lester, J. (2007). Early
prediction of student frustration. In A.C. Paiva, R.
Prada, & R.W. Picard (Eds.), Proceedings of the 2nd
international conference on Affective Computing and
Intelligent Interaction (ACII '07) (pp. 698-709).
Berlin, Heidelberg: Springer-Verlag.
Merrill, D.C., Reiser, B. J., Ranney, M., & Trafton, J. G.
(1992). Effective tutoring techniques: A comparison
of human tutors and intelligent tutoring systems. The
Journal of the Learning Sciences, 2(3), 277-305.
Raybourn, E.M. (2007). Applying simulation experience
design methods to creating serious game –based
adaptive training systems. Interacting with
Computers, 19, 206-214.
Schulze, K.G., Shelby, R.N., Treacy, D.J., Wintersgill,
M.C., VanLehn, K., & Gertner, A. (2000). Andes:
An intelligent tutor for classical physics. The Journal
of Electronic Publishing, 6(1).
Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC) 2011
2011 Paper No. 11010 Page 12 of 12
Shaffer, D.W., & Graesser, A. (2010). Using a
Quantitative Model of Participation in a Community
of Practice to Direct Automated Mentoring in an Ill-
Formed Domain. In C. Lynch, K. Ashley, T.
Mitrovic, V. Dimitrova, N. Pinkwart, & V. Aleven
(Eds.), Proceedings of the 4th International
Workshop on Intelligent Tutoring Systems and Ill-
Defined Domains held at the 10th International
Conference on Intelligent Tutoring Systems (ITS ‘10)
(pp. 61-68).
Simon, H.A. (1973). The structure of ill-structured
problems. Artificial Intelligence, 4, 181-204.
Sottilare, R., Holden, H., Goldberg, B., & Brawner, K.
(In Review). Considerations in the design of
modular, computer-based tutors for military
training. Submitted for publication in the Journal of
Military Psychology.
Thomas, R. C., & Milligan, C. D. (2004). Putting
teachers in the loop: tools for creating and
customizing simulations. Journal of Interactive
Media in Education (Designing and Developing for
the Disciplines Special Issue), 15.
U.S. Army Training and Doctrine Command (2000).
Objective Force capability. TRADOC Pam
525-66 Draft) Fort Monroe, VA.
VanLehn, K., Lynch, C., Schulze, K., Shapiro, J.A.,
Taylor, L., Treacy, D.,…Wintersgill, M.C. (2005).
The Andes physics tutoring system: Five years of
evaluations. In G. McCalla & C. K. Looi (Eds.),
Artificial Intelligence in Education (pp. 678-685).
Amsterdam: IOS Press.
Walker, E., Ogan, A., Aleven, V., & Jones, C. (2008).
Two Approaches for Providing Adaptive Support for
Discussion in an Ill-Defined Domain. In Proceedings
of the Workshop on Intelligent Tutoring Systems and
Ill-Defined Domains: Assessment and Feedback in
Ill-Defined Domains held at the 9th
International
Conference on Intelligent Tutoring Systems (ITS
’08), Montreal, Canada.
Weerasinghe, A., Mitrovic, A. & Martin, B. (2009).
Towards individualized dialogue support for ill-
defined domains. International Journal of Artificial
Intelligence in Education, 19, 357-379.
Zinsser, N., Perkins, L.D., Gervais, P.D. & Burbelo,
G.A. (2004). Military application of performance-
enhancement psychology. Military Review,
September-October, 62-65.
Zipperer, E., Klein, G., Fitzgerald, R., Kinnison, H., &
Graham, S. (2003). Training and Training
Technology Issues for Objective Force Warrior. ARI
Research Report 1809. U.S. Army Research Institute
for the Behavioral and Social Sciences: Alexandria,
VA.